NVIDIA used CES in Las Vegas on January 5, 2026, to introduce the Rubin platform, built around six new chips. Jensen Huang said Rubin is in full production, and NVIDIA claims up to ten times lower inference cost per token versus Blackwell. Reuters also reported Rubin delivers five times the AI computing power of its predecessor when serving chatbots and similar applications. Simply put, this directly targets the unit cost of using AI in live products.

Most leaders still frame AI as a training problem. Training is the expensive build phase, when a model learns from data and updates its internal weights. Inference is the usage phase, when the model answers real requests from staff or customers. Tokens are small chunks of text used to measure and price usage. When token cost and response time fall, the limiting factor stops being “can we run it” and becomes “where should we run it.”

That is why Rubin matters to strategy and operations. If the cost per response drops enough, teams stop rationing AI calls and start designing workflows that assume constant assistance. You can run stronger models more often, across more steps in a process, without turning every feature into a cost fight. This does not just improve quality, it changes distribution, because AI becomes cheap enough to include in standard plans and default experiences, not just premium tiers.

NVIDIA has been signaling this shift for over a year, and Rubin makes it explicit. The company is focusing on serving efficiency and token economics because that is where volume lies once AI is deployed. The long-run demand driver is not one training run, it is the daily stream of searches, drafts, summaries, support replies, and internal answers. NVIDIA’s own announcement frames Rubin around lower cost to accelerate mainstream adoption, not just bigger training.

The early 2025 market panic shows why this distinction matters. When DeepSeek launched, investors treated cheaper model economics as a direct threat to NVIDIA demand, and the market sold off hard. Reuters reported NVIDIA suffered a record one-day market value drop during that rout. The reaction assumed NVIDIA sells only training and that lower training cost would shrink total spend. It ignored the simpler driver, which is how many times models get used once they are good enough and cheap enough.

Inference is what scales in the real economy. Every time a customer support agent asks for a response, or a salesperson generates an email draft, or a manager summarizes a meeting, that is an inference bill. When those bills fall, usage expands because the friction is removed. That is how AI features move from “nice to have” to default. It is also how you get AI inside everyday apps like email in a way that feels instant, because response time is low enough to stay in the flow.

For leaders, the question is not whether Rubin is impressive. The question is whether your organization is ready to exploit cheaper inference when it arrives in late 2026, as Reuters reports on Rubin’s rollout context at CES. Many deployments stall today for two reasons: cost per use is too high, or response time is too slow, so adoption stays narrow. If that constraint eases, the bottleneck moves to workflow design, data access, and governance. That shift rewards teams that have already done the unglamorous work of ownership, policy, and measurement.

Here is the action plan.

First, measure cost per response and response time for the AI workflows you already run at scale, even if they are small.

Second, list the workflows you would expand if cost fell by an order of magnitude, based on NVIDIA’s own cost claim, and prioritize the ones tied directly to margin, speed, and retention.

Third, assign a single owner per workflow and require a clear rollout plan with guardrails so that cheaper compute leads to real deployment. Rubin changes the economics. Your advantage comes from being ready when the economics change.

Sources:

• NVIDIA, NVIDIA Kicks Off the Next Generation of AI With Rubin — nvidianews.nvidia.com
• Reuters, Nvidia CEO Huang says next generation of chips is in full production — reuters.com
• Reuters, DeepSeek sparks AI stock selloff; Nvidia posts record market cap loss — reuters.com